Determine whether SCAFFOLD instability is intrinsic

Determine whether the collapses observed for SCAFFOLD under the reported federated-learning setting are intrinsic by conducting a tuned comparison that selects the server step size on seeds 1–5.

Background

The supplementary experiments report that SCAFFOLD, evaluated at the shared learning rate without the hyperparameter selection applied to the other tuned methods, collapsed on all five CIFAR-10 reporting seeds and on four of five FEMNIST reporting seeds. The appendix attributes this behavior tentatively to the fragility of control variates when clients participate infrequently and train for many local steps, but does not establish whether the collapses are inherent to SCAFFOLD or result from the untuned server step size. A comparison that tunes this parameter on the selection seeds would resolve that uncertainty.